๋ณธ๋ฌธ ๋ฐ”๋กœ๊ฐ€๊ธฐ

IN DEPTH CAKE/ML-WIKI

(3)
Inductive Bias, ๊ทธ๋ฆฌ๊ณ  Vision Transformer (ViT) ๋“ค์–ด๊ฐ€๋Š” ๋ง Transformer๋Š” (CNN๋ณด๋‹ค) Inductive Bias๊ฐ€ ์•ฝํ•œ ๋„คํŠธ์›Œํฌ๋กœ, general-purpose ๋„คํŠธ์›Œํฌ์˜ ์ƒˆ๋กœ์šด ์ง€ํ‰์„ ์—ฐ ๊ตฌ์กฐ๋กœ ํ‰๊ฐ€๋ฐ›์Šต๋‹ˆ๋‹ค. Inductive Bias๊ฐ€ ์ ๋‹ค๋Š” ๊ฒƒ์€ ์–‘๋‚ ์˜ ๊ฒ€์ธ๋ฐ, ์ด๋ฅผ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” Inductive Bias๊ฐ€ ๋ฌด์—‡์ธ์ง€ ๊ทธ๋ฆฌ๊ณ  Inductive Bias๊ฐ€ ํ•™์Šต์— ๋ผ์น˜๋Š” ์˜ํ–ฅ์„ ์ดํ•ดํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ณธ ๊ธ€์—์„œ๋Š” ์ถ”์ƒํ™”๋œ ํ˜•ํƒœ๋กœ Inductive Bias๋ฅผ ์„ค๋ช…ํ•ด๋ณด๋ ค ํ•ฉ๋‹ˆ๋‹ค. In computer vision, there has recently been a surge of interest in end-to-end Transformers, prompting efforts to replace hand-wired features or i..
<ML๋…ผ๋ฌธ> ํด๋ž˜์Šค ๋ถˆ๊ท ํ˜• ๋ฌธ์ œ Cui et al. "Class-Balanced Loss Based on Effective Number of Samples" (CVPR 2019) TL;DR Class imbalance ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด์„œ ๋ฐ์ดํ„ฐ์…‹ ๊ฐ ํด๋ž˜์Šค์˜ ์œ ํšจ ๋ฐ์ดํ„ฐ ์ˆ˜๋ฅผ ์ •์˜ํ•˜๊ณ  ์ด๋ฅผ ํ™œ์šฉํ•œ re-weighting๊ธฐ๋ฐ˜ Class Balance Loss ๊ธฐ๋ฒ• ์ œ์•ˆ. ๋ฌด์Šจ ๋ฌธ์ œ๋ฅผ ํ’€๊ณ  ์žˆ๋‚˜? ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ ํ•™์Šต์— ์‚ฌ์šฉ๋˜๋Š” ์ผ๋ฐ˜์ ์ธ ๋ฐ์ดํ„ฐ ์…‹ (์˜ˆ๋ฅผ๋“ค์–ด CIFAR-10, 100, ImageNet ๋“ฑ)์ด ํด๋ž˜์Šค ๋ผ๋ฒจ ๋ถ„ํฌ๊ฐ€ ๊ท ์ผํ•œ ๊ฒƒ๊ณผ ๋‹ฌ๋ฆฌ, ์‹ค์ œ ์ƒํ™ฉ์—์„œ๋Š” ๋ชจ๋“  ํด๋ž˜์Šค์˜ ๋ฐ์ดํ„ฐ ์ˆ˜๊ฐ€ ๊ท ์ผํ•˜๊ฒŒ ์ˆ˜์ง‘๋˜์ง€ ์•Š๋Š”, Long Tail ํ˜„์ƒ์ด ๋ฐœ์ƒํ•œ๋‹ค. ์—ฌ๊ธฐ์„œ Long Tail์ด๋ผ๊ณ ํ•จ์€, ๊ฐ ํ•™์Šต ๋ฐ์ดํ„ฐ ์˜ ํด๋ž˜์Šค ๋ณ„ ์ƒ˜ํ”Œ ์ˆ˜์— ๋Œ€ํ•œ ๋ถ„ํฌ๋ฅผ ๊ทธ๋ ธ์„ ๋•Œ ์•„๋ž˜ ๊ทธ๋ฆผ๊ณผ ๊ฐ™์ด, ์†Œ์ˆ˜์˜ ํด๋ž˜์Šค์— ๋Œ€ํ•ด์„œ ๋ฐ์ดํ„ฐ ์ƒ˜ํ”Œ ์ˆ˜๊ฐ€ ๋งŽ์€ ๋ฐ ๋ฐ˜ํ•ด (Head) ๋‹ค์ˆ˜์˜ ํด๋ž˜์Šค์—์„œ ๊ธฐ๋Œ€์น˜ ์ดํ•˜์˜ ์ƒ˜ํ”Œ ์ˆ˜๋ฅผ ๊ฐ–๋Š” (..
<ML๋…ผ๋ฌธ> CVAE์— ๋Œ€ํ•˜์—ฌ (feat. ๋ˆ„๊ฐ€ ์ง„์งœ CVAE์ธ๊ฐ€? ํ•˜๋‚˜์˜ ์ด๋ฆ„, ๋‘ ๊ฐœ์˜ ๊ธฐ๋ฒ•) ๋ณธ ๊ธ€์€ ๋…ผ๋ฌธ์˜ ์ƒ์„ธ ๋‚ด์šฉ์„ ๋‹ค๋ฃจ๋Š” ํฌ์ŠคํŒ…์€ ์•„๋‹ˆ์—์š”. ๋‹ค๋งŒ, ๋‘ ๊ฐœ์˜ ๋…ผ๋ฌธ์ด ํ•˜๋‚˜์˜ ์ด๋ฆ„์œผ๋กœ ๋ถˆ๋ฆฌ๊ณ  ์žˆ๊ธธ๋ž˜, '์ด์— ๋Œ€ํ•œ ํ˜ผ์„ ์„ ์ •๋ฆฌํ•˜๋Š” ๊ธ€์„ ์จ๋ณด์ž'ํ•˜๋Š” ๋งˆ์Œ์œผ๋กœ ๊ธ€์„ ์ผ์Šต๋‹ˆ๋‹ค. ๋ฌผ๋ก  ๊ฐ„๋žตํ•˜๊ฒŒ ๊ฐ๊ฐ์˜ ๋…ผ๋ฌธ ๋‚ด์šฉ์„ ๋‹ค๋ฃจ๊ณ  ์žˆ๊ธฐ๋Š” ํ•ฉ๋‹ˆ๋‹ค๋งŒ, ์ž์„ธํ•œ ๋‚ด์šฉ์€ ๊ฐ ๋…ผ๋ฌธ ๋‚ด์šฉ์„ ์„ค๋ช…ํ•˜๋Š” ๊ธ€์„ ์ฐธ๊ณ ํ•ด์ฃผ์„ธ์š”. ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์•ˆ๋…•ํ•˜์„ธ์š”. ์˜ค๋Š˜์€ ๊ฐœ์ธ์ ์œผ๋กœ ํฅ๋ฏธ๋กœ์› ๋˜ ํ˜„์ƒ์— ๋Œ€ํ•ด ์ ์–ด๋ณผ๊นŒ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ณธ๋ฌธ์„ ์ฝ๊ณ  ๊ณ„์‹  99.9 % ์˜ ๋ถ„๋“ค์€ "CVAE"๋ผ๋Š” ํ‚ค์›Œ๋“œ๋ฅผ ๊ฒ€์ƒ‰ํ•˜์…จ์„ ๊ฒƒ ๊ฐ™์•„์š”. ์—ฌ๋Ÿฌ๋ถ„์ด ์ด ๊ธ€์„ ํด๋ฆญํ•˜์‹ค ๋•Œ ์ƒ๊ฐํ•˜์‹  CVAE๋Š” ์–ด๋–ค ๋…€์„์ธ๊ฐ€์š”? ์งˆ๋ฌธ์ด ์ด์ƒํ•˜์ฃ ? ์ œ๊ฐ€ ์˜ค๋Š˜ ์ด ๊ธ€์„ ์“ฐ๊ธฐ๋กœ ๋งˆ์Œ๋จน์€ ๋ฐ๋Š” ์ด์œ ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ œ๊ฐ€ ๊ฒ€์ƒ‰์„ ํ•˜๋‹ค๊ฐ€ ๋ฐœ๊ฒฌํ•œ ํ˜„์ƒ์ด ์žˆ๋Š”๋ฐ, ๋ฐ”๋กœ ๋‘ ๊ฐœ์˜ (์œ ๊ด€ํ•˜์ง€๋งŒ ์„œ๋กœ ๋‹ค๋ฅธ) ๋…ผ๋ฌธ์ด ..